Connect the agent to what customers can already see, and nothing else. It reads published articles and their metadata, keeps its index in step with every publish and unpublish, and cites article versions in its answers. It may create drafts for owners to review. It never publishes, edits live articles or reads internal-only content when answering customers.
What does the connection involve, and who owns it?
A help centre is a content system: articles, categories, versions, owners, review dates, regional variants and often search analytics. Connecting an agent to it makes the published content the agent's source for product questions. The connection decides three things: what the agent reads, how quickly it learns about changes, and whether it can change anything.
The content or knowledge team owns the help centre. Article owners, usually in product, own individual articles. Security owns the agent's identity. Support owns the use of the agent. The same principle runs through our guide to the AI support agent that knows when to say no: narrow sources make for answers you can defend.
What are the steps, and at what depth?
- Person, content lead: creates a dedicated identity for the agent in the help centre system, with a named owner. Not an editor account.
- Person, content lead: grants read access to published articles in the regions and languages in scope. Depth 1.
- Person, content lead: grants read access to metadata: owner, last reviewed date, region, product, version ID. The agent needs these to choose and cite correctly.
- Person, security: keeps internal articles out of the customer-facing source set. If staff also use an agent on internal knowledge, it is a separate source with separate access.
- Person, systems team: wires change events. Publishing, updating and unpublishing an article updates the agent's index promptly. An unpublished article must stop being cited at once.
- Agent, depth 1: reads search analytics, where available, to see what customers looked for and did not find.
- Agent, depth 2: creates draft articles and suggested edits in a drafts area, assigned to the article owner.
- Person, article owner: reviews and publishes. Publishing, editing live articles and deleting are not in the agent's role.
- Person, security: tests revocation and confirms the agent stops answering from the help centre when access is removed, rather than answering from a stale copy.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Published articles | Reads in-scope content | 1 | Source scope |
| Metadata | Reads owner, date, version | 1 | Nothing |
| Internal articles | Nothing in customer use | none | Separate access decision |
| Change events | Updates its index | 1 | Wiring and monitoring |
| Search analytics | Reads failed searches | 1 | Nothing |
| Drafts | Creates drafts and suggestions | 2 | Owner reviews |
| Publish, edit, delete | Nothing | none | Article owners |
Which controls matter, and what evidence does each leave?
Published only. Drafts and archived articles are never customer sources. Evidence: the status of every article cited, checked at answer time.
Version cited per answer. If an article was wrong, you can find every customer who was told it. Evidence: article ID and version on each answer.
Index freshness is monitored. The gap between a publish event and the agent's index update is measured. Evidence: freshness logs and alerts.
Drafts are clearly drafts. The agent's suggestions sit in a drafts area with the agent named as author. Evidence: authorship and the owner's decision.
No publishing capability. Evidence: the role's permission list and refused verdicts for any attempt.
What should you measure?
- Index lag: time from publish or unpublish to the agent's index reflecting it.
- Answers citing articles past their review date.
- Drafts created and published, and the time owners take to review.
- Answers citing unpublished articles. The target is none; one is an incident.
- Customer searches with no result, as an input to finding gaps in the help centre with an AI agent.
Where does it usually go wrong?
The agent is given an editor role. It is the easiest existing role to reuse, and it can publish. Build a read and draft role instead.
Unpublishing does not reach the agent. A withdrawn article about a discontinued product keeps being cited for weeks. Wire the unpublish event, not just publish.
Internal and external content mix. An internal troubleshooting note leaks into a customer answer, along with details customers should not see. Keep the sources apart.
Regional variants collapse. The connection pulls one version of each article and loses the regional ones. Carry the region through.
Embedded content is invisible. Tables, videos and images often hold the actual instructions, such as fee tables or step-by-step screenshots. If the connection extracts only body text, the agent misses them and answers from the surrounding sentences. Check what the connection actually extracts from a sample of your most-used articles before going live.
Answers outrun the content. The agent becomes more trusted than the articles it reads. Answering product questions from the help centre with an AI agent covers how it should behave when an article is silent, and can an AI agent help a vulnerable customer? covers when an article is not enough.
Where Pryme fits
In Pryme, the help centre is connected as a named source through the Pryme Context Protocol, with each article's version carried into the answer and the filed record. Draft creation is a separate capability at depth 2; publishing is never part of the role, so any attempt is refused and recorded.
Questions people ask
Should the agent read the help centre live or from a copy?
Either works if the copy is kept current by change events and its freshness is measured. Live reads avoid staleness but depend on the help centre being available.
Can the agent tell us which articles customers struggle with?
Yes. Combining failed searches, routed questions and repeat contacts gives article owners a clear list.
Does a supervisor care what the help centre said?
It can matter. What customers were told, and when, is evidence in a conduct review, which is one reason to keep versions. Can an AI agent help prepare for a regulatory inspection? covers assembling that kind of evidence.